Written by an AI agent working in this repository. Synthetic data, not tax advice.
I wrote a made-up October bank statement for a solo therapist (fixtures/therapist-checking.csv) and ran it through ai-bookkeeper's importer before writing any rules. Of 14 lines, 6 got an account. Headway and Alma payouts, a CE course, clinical supervision, APA dues, a Psychology Today listing and the state board renewal did not.
I added built-in rules for panel payouts (sales), CE, supervision and association dues (education and dues), directory listings (advertising) and the board licence (taxes and licenses). Now 13 of 14 get a suggestion. The Amazon line stays blank on purpose: an office chair may be equipment to depreciate and test kits are supplies, so the therapist decides. The $500 Venmo cash-out is filed as an owner draw, which keeps the month's profit at $2,713.56 rather than $2,213.56.
A therapist's bank statement can hint at clients, so the page makes a point of the tool running locally in the browser with nothing uploaded.
The rules came from company names and this fixture's descriptions, not a real practice's export, and were tuned on this one fixture, so "13 of 14" is the fixture passing, not an accuracy figure. A real export is still what the project needs.